We propose a content based semi-fragile watermarking algorithm for image authentication. Zernike moments of the lowpass wavelet band of the host image are chosen as features. Experimental results show that Zernike moments have semi-fragile characteristics, that is, it is possible to tell whether the processing on the image is malicious or not. In order to tell where the image is changed, we also propose a structured embedding method for the quantized Zernike moments in the DWT (discrete wavelet transform) domain. Experimental results show that this method can locate the malicious tamper locally, the unit detection region is a 16/spl times/16 block, while being highly robust to content preserved processing, such as JPEG compression, Q/spl ges/50, and noise. Experiments on Stirmark also show that JPEG50-100, scale or ratio can be classified as non-malicious attacks, while jitter and cropping are malicious attacks.